While putting #196 together I noticed this comment:

Is this actually true for the code here? This doesn't look like a lazy pattern match to me; forcing the first cons of reverse ys will force the linear (in the length of zs) amount of computation as part of the STM transaction AFAICT. If so, it probably makes sense to either fix this (but it isn't high priority as it isn't used by IO) or remove the comment.

Also see the upstream readTBQueue for comparison: https://github.com/haskell/stm/blob/07df1050ab5d45244f718eee8a512bb18e7066f6/Control/Concurrent/STM/TBQueue.hs#L133-L140